Académique Documents
Professionnel Documents
Culture Documents
Abstract
When computing time-domain impedance profile from measured S-parameters, we face two types of problems. First is caused by low quality of sampled S-parameters, such as insufficient resolution, band-limiting,
noise etc. These limitations create certain numerical challenges, such as response aliasing, ringing and noncausality. The computed impedance becomes sensitive to termination conditions set up at the far end of
the connector. The second problem is the fact that even with the ideal data, the computed impedance profile does not show the correct characteristic impedance of each section of the connector, because multiple
reflection mask the true value of the characteristic impedance of the distant sections. In the paper, we
address both problems: first, by post-processing inverse discrete Fourier transform (IDFT) results, or by
vector fitting. Second is addressed by modeling the device under test (DUT) as a cascade connection of
multiple small transmission lines and applying time domain gating and peeling, to calculate the length and
impedance of each transmission line, so that the impedance profile of the combined cascaded system is
equivalent to that of the DUT.
Biography
Dr. Vladimir Dmitriev-Zdorov is a Principal Engineer in the System Design Division at Mentor Graphics
Corporation. He has developed a number of advanced models and novel simulation methods used in the
companys products. Currently his work includes development of the efficient methods of circuit/system
simulation in time and frequency domain, transformation and analysis of the multi-port systems, statistical
& time domain analysis of SERDES links. He received Ph. D and D. Sc degrees (1986, 1998) based on his
work on circuit and system simulation methods. The results of his work have been published in numerous
papers, conference proceedings, and a monograph.
Dr. Kaviyesh Doshi received his Ph. D in electrical engineering from University of California, Santa
Barbara in 2008.
He joined Teledyne LeCroy, a manufacturer of high-performance measurement instruments located
in Chestnut Ridge, New York in 2008 as a Research & Development Engineer. At Teledyne LeCroy, he
has been involved in the design and development of SPARQ - a time-domain reflectometry (TDR) based
instrument for measuring S-parameters.
Dr. Doshi has one patent and many others applied for in the area of de-embedding and signal processing
for time-domain network analysis.
Pete Pupalaikis received the B.S. degree in electrical engineering from Rutgers University, New Brunswick,
New Jersey in 1988.
He has worked at Teledyne LeCroy (formerly LeCroy Corporation) located in Chestnut Ridge, New
York since 1995 and currently manages integrated-circuit development, signal processing and intellectual
property as Vice President of Technology Development. Prior to LeCroy, he served in the United States
Army and as a consultant in embedded systems design.
Mr. Pupalaikis has thirty two patents and has published numerous papers in the area of measurement
instrument design and is a member of Tau Beta Pi, Eta Kappa Nu and the IEEE signal processing, instrumentation, solid-state circuits and microwave societies. In 2013 he was elevated to IEEE Fellow for
contributions to high-speed waveform digitizing instruments.
Introduction
HE TIME-DOMAIN IMPEDANCE PROFILE of cables and connectors has remained for decades an
important indicator of their characteristics such as characteristic impedance and delay, or transmission problems that include reflections caused by impedance discontinuities, attenuation, skew
and some others. Historically, time-domain reflectometry (TDR) impedance first was a measurement
technique. Later on, when operational frequencies advanced, the need for detailed electrical simulation
necessitated full S-parameter measurements, and the connector vendors supplied either TDR results and
S-parameters, or more frequently - only S-parameters. Today, TDR impedance characteristics remain of
great importance for designers, but more and more often these should be found directly from S-parameters
by computations. To save processing time and disk space, the sampling of measured S-parameter data
is often made barely enough to correctly represent the connectors delay, and the upper frequency comes
short to what is required by the Nyquist criteria for a given data rate. In TDR impedance computations,
we deal with two times the connector delay, which makes given S-parameters even more under-sampled
considering this particular purpose. The latter creates some computational challenges which we are going
to address in the paper. The second issue, perhaps even of greater importance, is the fact that what we
observe in the impedance plot may not be the actual characteristic impedance of the transmission line, if it
is masked by a segment with different characteristics.
A commonly used method to obtain the impedance profile from S-parameters is by using inverse discrete Fourier transform (IDFT) of the return loss. This method does not take into account the multiple
reflections due to impedance discontinuities. These impedance discontinuities introduce multiple reflections, which act as secondary sources. If not accounted for, these multiple reflections can give incorrect
result for impedance profile. In this paper we propose an alternate way of computing the impedance profile
by modeling the line as that obtained from the cascade connection of small transmission line structures.
The method involves calculating the impedance for small section at a time and then undoing the effects of
this small section before proceeding to calculate the impedance for the following sections. As a by-product
of the algorithm, we describe the assumption involved in calculating the impedance profile by performing
the IDFT of the return loss. It turns out that for situations where the characteristic impedance of the line
is close to the reference impedance, the approximate method of using the IDFT does not introduce large
errors and can be used. Nevertheless appropriate care needs to be taken to compute the IDFT of return loss
to account for the discrete nature of the S-parameters, band-limited data and the frequency spacing of the
data.
The paper is organized as follows. First, we show some basic relations between characteristic impedance,
voltage transfer functions and S-parameters. Then, we consider two methods of converting frequencydomain characterization into time domain: by inverse Fourier transformation of the sampled dependence
and then by rational fitting represented in a form of rational fraction expansion (RFE). The two approaches
are applied to the same connector model. We show that different termination conditions on the far end may
affect the initial portion of the impedance plot if the data is not perfect. In the next half we describe how
the impedance profile can be recovered by modeling the line as that obtained from a cascade connection
of small transmission line structures. An algorithm is provided which can be easily incorporated into a
computer program.
1
Z (t)
= Z0
v1 (t)
E (t) v1 (t)
(1)
Expression (1) gives a practical way of finding an impedance profile for any linear model, not necessary
the transmission line. Lets consider S-parameters, as in Figure 2. To find an impedance profile for a
certain port, we need to find the ports input voltage and then apply (1). Lets first define the port voltage
transfer function as:
Kij (s) =
Vi (s)
Ej (s)
(2)
Note that (2) assumes that the voltage stimulus is applied only to port j while the voltage is measured
at port i, including the case i = j, and that all reference resistances shown in Figure 2 are port normalizing
impedances.
It is easy to prove a couple of simple relations between the port voltages found in the topology of Figure
2 and the return and insertion loss parameters. In particular, if the stimulus and the port voltage refer to the
same port i, we can express the return loss as:
sii =
2Vi (s)
1
Ei (s)
(3)
If the stimulus is applied to the port j and the port voltage is measured at the port i, we can express the
mutual port coefficient:
2Vi r0j
sij =
(4)
Ej r0i
From (2) and (3) it follows that:
1
(1 + sii )
2
If the voltage stimulus is a Dirac impulse (i.e. Ej (t) = E0 (t)), the port voltage becomes:
Kii (s) =
E0
(1 + sii )
(5)
2
By converting (5) to the time-domain and integrating, we find the voltage response on the unit step stimulus,
and then the port impedance from (1). Figure 2 implies that the S-parameters describe a single-ended
model. If we want to find differential or common impedance for the mixed-mode model, we still can use
the same approach, with the only difference that sii in (5) is replaced by the port differential- or commonmode return-loss, and the normalizing impedance in (1) should be doubled or halved respectively. In
particular, if the differential port is made from the pair of single ended ports (k, m) and these single ended
ports have the same normalizing impedance r0 (as required by the Touchstone 2.0 format specification),
then (5) becomes one of the following:
]
[
vdiff (t)
1
E0
(6)
1 + (skk skm smk + smm ) , zdiff (t) = 2Z0
Vdiff (s) =
2
2
E (t) vdiff (t)
Vi (s) =
(a) Un-zoomed
[
]
E0
1
1 + (skk + skm + smk + smm ) , zcomm (t)
2
2
1
vcomm (t)
Z0
(7)
2 E (t) vcomm (t)
What is different if some ports, for example far end ports of the connector model, are not terminated by
the reference impedance? In this case we have to exclude these ports by considering their actual termination
conditions. Such transformation affects parameters of all ports, but can be easily computed directly in
frequency domain for any sampled data. To mitigate the effect of band limiting, it is reasonable to choose
the stepwise stimulus with a finite rise time, which is equivalent to high frequency filtering. To do that,
time domain integration of the voltage Dirac response is slightly modified and represented by two integrals,
one for the linear ramp and the other for the constant part of the stimulus.
Am
1 + sm
m=1
(8)
From (8), the time-domain response can be found directly. For example, unit step response becomes:
[
]
M
a (t) = H0
Am em t for t 0
(9)
M
m=1
where H0 = H + m=1 Am . We are going to illustrate the two methods in parallel and show how data
sampling affects the accuracy of time responses. We take a typical 4-port connector model, with about 3.5
4
(a) Un-zoomed
(c) Trajectory Plot showing Imaginary vs. Real Part with Different
Sampling
(a) Time-domain
(b)
Frequency-domain
(Real Part)
(a) 2 MHz
(b) 50 MHz
(c) 70 MHz
Figure 11: IDFT with Post-processing with Different Sample Frequency Resolution
Figure 12: RFE Comparison with Original Data for Varying Resolution
Red is RFE Approximation. Blue is Original Data.
can be re-sampled with arbitrary fine step up to arbitrary large frequency. However, as we understand, the
RFE was created at some point from measured or simulated data which could be under-sampled or have
errors. To a certain extent, the RFE may improve data quality, but if it was non-causal or poorly sampled, it
may add an unintended behavior, compare Figure 12(a) and Figure 12(b). Once created, the RFE cannot
be improved by interpolation because it is already an analytical representation. Fitting/building an RFE
should preferably be made to the complete non-reduced matrix, where the complexity of every component
in term of poles/residues is typically less than of reduced matrix or the voltage transfer function.
To find an impedance profile, we need to find the port voltage in time domain. As follows from (5-7),
voltage transfer functions in RFE form can be found directly from an RFE representation of S-parameters,
but only if all other ports are terminated by their normalizing impedances. To evaluate the error of TDR
computations, wed like to have them done with different termination conditions, as we did with the IDFT.
To make it possible, we either need to re-sample the RFE and transform S-matrix for every frequency point,
or find the resulting transfer function in RFE form directly. These are possible scenarios:
1. Re-sampling, matrix conversion, re-fitting or IDFT
(a) Re-sample RFE defined S-parameters with sufficiently fine resolution, to allow more complexity than can be seen for each matrix component.
(b) Find sampled dependence of the required return loss characteristic with appropriate termination
applied
(c) Re-fit found dependence
(d) Convert new fit into time response, then impedance plot
(e) A version with the IDFT: instead of (c,d), apply the IDFT to sampled dependence
2. Finding new poles from a state-space representation (SSR)
(a) Convert an original RFE matrix into state space equations.
(b) Apply termination conditions, update matrices in state space representation
(c) Convert the new SSR into a scalar RFE for the return loss parameter
(d) Convert the RFE into the time response and impedance plot
In the first approach, we found that performing the IDFT after re-sampling the RFE by 10 MHz allows
more accurate impedance computation than with re-fitting, as the deviation between different cases does
not exceed 0.3-0.5 % versus 0.7 %, (see Figure 13(a) and Figure 13(b)). This is because precise refitting
is difficult for the transfer function, due to its higher complexity.
The second approach is based on the fact that RFE can be converted directly, without re-sampling and
refitting. It is known that RFE is equivalent to a certain state space representation in (10) where X CP is
a vector of state variables; U CN is an input and Y CN is an output vector:
( ) (
)( )
( ) (
)(
)
x
x
0
sI
X
=
or
=
(10)
y
u
Y
U
We assume that Y and U have identical size and therefore input /output relationships can be described
by a square matrix , as in case of S-parameters:
10
[
]
Y = ( sI)1 + U = H (s) U
(11)
The relationships between the matrices , , , and poles/residues are known (see e.g. [4]) so that
RFE in form (8) can be built from these matrices. The poles then become eigenvalues of , and the
coefficients are expressed via and . Conversely, if the RFE exists for all components of the transfer
matrix H (s), the matrices , , and can be easily found, too.
Our task is to find the RFE of the reduced transfer matrix, in assumption that some ports of the Sparameter model are eliminated because of certain termination conditions. Let the original equations be:
)
(
)(
) (
B1
S11 S12
A1
=
(12)
B2
S21 S22
A2
The termination conditions applied to the ports collected in the second sub-vector:
A2 = RB2
(13)
zi z0i
The components of the diagonal matrix R are reflection coefficients of the terminator, i =
,
zi + z0i
which always exist if the termination is passive; z0i , zi are port normalizing and terminating impedances
respectively. Since the reflected wave B is an output, and the incident wave A is an input, we can rewrite
(10) as:
X
sI 1 2
0
Y1 = 1
11 12 U1
(14)
U2
2
21 22
Y2
In (14), U2 = RY2 . From the last equation: Y2 = 2 X+21 U1 +22 RY2 we get Y2 = (I22 22 R)1 2 X+
(I22 22 R)1 21 U1 and U2 = R (I22 22 R)1 2 X + R (I22 22 R)1 21 U1 .
Then, (14) reduces to:
(
0
Y1
(
=
)(
X
U1
)
(15)
If the termination impedance is not frequency dependent (e.g. resistive, open or grounded) then only
the upper left block changes with frequency, and (15) and (10) have similar structure. Finally, from (15) we
get the new state-equation matrices and build a new RFE that describes the reduced model with some ports
terminated, grounded or disconnected. Then from (5-7), we find the differential voltage transfer function,
voltage waveform and an impedance profile. With grounded or disconnected far end, the converted RFE
contains much more poles (real ones and complex conjugate pairs), 404 total, than in case of the matched
termination (340). The TDR impedance plot is shown in Figure 13(c). Here we observe better agreement
between the curves, with a difference of only about 0.1%.
In the above transformations, we assumed that the RFE represents S-parameters with no error, therefore all discrepancies in TDR impedance plots were caused by the transformations. With non-ideal RFE,
TDR impedance plots found with different termination conditions would manifest a certain non-removable
difference. The root of this difference is inaccurate representation of the delay operator embedded into
the voltage transfer function. This makes a small part of the reflected power to appear as non-reflected
and start earlier in time than it should.
11
1 + a
1
a1
b1
1
b1
a2 1
a2
z1
b2
b2
1+
Zc Z0
.
Zc + Z0
(16)
For the ideal, lossless transmission line with no frequency dependent delay, the delay of the line T is
incorporated in z as:
z = ej2fT .
(17)
The transmission line model shown in Figure 14 can be thought of as composed of cascade connection
of three different two-port networks as shown in Figure 15. Describing from the left, the first dotted line box
corresponds to a network that models the impedance change, i.e. the wave will see a change of impedance
from Z0 to Zc and this is reflected by an S11 of value . Due to this impedance change, part of the wave
will be reflected and rest of the wave will be transmitted through. The next two-port section is shown by
the solid box. Here the characteristic impedance is Zc and the wave does not see any impedance change.
As a result the wave gets transmitted through. The through wave is delayed by the S21 = z1 . The third
section is again an impedance transformer what transforms the impedance the wave sees - from Zc to Z0 .
The combined effects of all the impedance transformations can be described by the S-parameters of the
transmission line model of Figure 14:
)
(
1
(1 z2 ) (1 2 ) z1
.
(18)
2
1
2
(1 ) z
(1 z )
1 z2 2
Some more reality can be added by incorporating a loss and group delay that is a function of frequency
G(f) and D(f), respectively. For the line that is lossy or contains frequency dependent loss or delay, we use
the following equation:
12
Z0 Zc
a1
1+
Zc Z0
1 a
2
z1
b1
z1
b2
1+
Zc Zc
xk
1 + 0
0
z1
1 0
vr0 k
1 0 1 + 1
0
0
z1
vfM k
vf2 k vfM1 k
vf1 k
vf0 k
1
0
1 + 0 1 1
vr1 k
1 1
z1
z1
1 + M1
M1
z1
M1
1 + 1
1 M1
vr2 k vrM1 k
1 1
M1
M1
z1
1 + M1
vrM k
z = G(f)ej2f(T+D(f))
(19)
The model shown in Figure 14 forms the basic unit i.e. in the impedance profile model, we consider
the line to consist of several such small transmission line sections cascaded together . This model is shown
in Figure 16. For this model, we will first generate the equations that allow the determination of forward
and reverse waves at various times and locations along the line, knowing the and z of each section. Then
we will address the inverse problem of determining the and z for each individual section from the system
S-parameters.
To simplify things, we start by simplifying the interface between two transmission line sections:
vfm k
1 m1
z1
m1
1 + m
m
m1
z1
z1
m
z1
1 m
1 + m1
vrm k
A=
0
0
0
m1
1
0
(1 + m1 )
(1 m1 ) 0
1
m1
0
0
m
1
0
0 (1 + m )
0
0
0
0
0
(
)T
v = a1 b1 f r b2 a2
(
x=
1 0 0 0 0 0
0 0 0 0 0 1
0
0
0
0
0
0
0 (1 m )
1
m
0
1
(20)
(21)
)T
(22)
Such that:
A1 x = v
(23)
where the first column in v represents the waves present at all of the nodes defined in v when the
network is driven only with unity incident waves from the left and the second column in v represents the
waves present at all of the nodes defined in v when the network is driven only with unity incident waves
from the right.
Therefore, the equivalent S-parameters of this network can be written as:
Sinterface
( (
) )
(
)
A1 x 1,0 A1 x 1,1
(
)
=
= ( 1 )
A x 4,0 A1 x 4,1
m1 m
m m1 + m m1 1
m m1 + m1 m 1
m m1
m m1 1
(24)
Next, we obtain the values of the middle node f and r which represent the forward and reverse waves
at the middle of the interface as a function of the incident waves on the network:
(
) )(
( ) ( ( 1 )
)
A x 2,0 A1 x 2,1
f
a
1
(
)
= ( 1 )
=
r
a2
A x 3,0 A1 x 3,1
14
m1 1
m1 (m 1)
m (m1 1)
m 1
m m1 1
)
(
a1
a2
)
(25)
xk
F0 k P0 21
z1
P0 22
F1 k P1 21
P1 11
P1 22
z1
R0 k
z1
F2 k P2 21
PM1 21
P2 11
z1
FM k
PM1 22 PM 11
z1
z1
P1 12 R1 k
P2 12
PM1 12 RM1 k
PM 12 RM k
Zm Z0
Zm +Z0
and therefore,
(26)
Where:
m =
Zm Zm1
m m1
=
Zm + Zm1
1 m m1
(27)
(27) is intuitive because it states that the S-parameters of the interface between two lines is given by the
mismatch between the two lines, not the mismatch between the impedance of the lines and the reference
impedance Z0 . In other words, cascaded lines that have the same characteristic impedance do not cause any
reflections or degradation of the forward propagating wave at the interfaces because the lines are matched
to each other (the only reflections occur getting the signal into this line in the first place when the driving
impedance is 50 .)
The Solution for the Waves Along the Line
For the moment, lets disregard the waves at the interfaces of the lines and instead concentrate on some
nodes that we can work with easily. In this spirit, we redraw the model shown in Figure 16 using the Sparameters for Sinterface as defined in (26) to obtain a new diagram as shown in Figure 18. In Figure 18, we
define Pm to be the S-parameters of the interface between sections m 1 and m. Also, the nodes Fn k and
Rn k represent the waves at the outputs of the forward and reverse delay taps, respectively. Note that Fn k
and Rn k dont really exist and will need to be converted to the actual forward and reverse going waves at
the transmission line section interfaces. Note also that there is a small problem at the ends of the line. The
problem is that for a series of cascade lines, there is no defined before the beginning of the line or after
the end of the line. This is handled by defining rho1 and rhoM as the source and terminating impedance.
Usually, this is set to zero to indicate that the system starts as being driven through an impedance equal to
the reference impedance Z0 and ends in a termination of Z0 , but this is arbitrary.
Now, lets write the equations for the Fm k and Rm k . We write these by going no further from the node
being defined to another defined node. We therefore get:
Fm = z1 Pm1 21 Fm1 + z1 Pm1 22 Rm1
(28)
(29)
xl 0 0 0
0 0 0 xr
)T
1
0
z1 P021
0
..
.
0
1
z1 P022
0
..
.
0
z1 P111
1
0
..
.
0
z1 P112
0
1
..
.
..
.
0
0
0
0
..
.
0
0
0
0
..
.
0
0
0
0
..
.
0
0
0
0
..
.
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
z1 PM1 21
0
0
1
z1 PM1 22
0
0
z1 PM11
1
0
0
z1 PM12
0
1
F0 R0 F1 R1
FM1 RM1 FM RM
)T
(30)
Thus, for M sections, we have a system matrix A of order 2(M + 1), a nodes vector with 2(M + 1)
elements, where M + 1 nodes are forward and M + 1 nodes are reverse. To make sense of this, consider
that if we have no sections, we would have two nodes, one forward and one reverse, but for a single section
we have four nodes: A forward and reverse wave at each interface to the section. With two sections, we
have six nodes, etc. The system matrix is thus defined as the 2(M + 1) 2(M + 1) element identity matrix
with, for m {0 . . . M 1}, the following additions:
A2(m+1),2m = z1 Pm 21
(31)
A2(m+1),2m+1 = z1 Pm 22
(32)
A2m+1,2(m+1) = z1 Pm+1 11
(33)
A2m+1,2(m+1)+1 = z1 Pm+1 12
(34)
With this, we now have all of the tools necessary to solve for the waves along a transmission line
modeled as described up to now. Given a sample rate Fs for the waves and number of sections M of length
1
and the characteristics of each section m {0 . . . M 1}, the values that define each section,
T = 2Fs
namely Zm , Gm (f) and Dm (f), and a frequency point n {0 . . . N} where fn = Nn Fs
:
2
1. Calculate the values of m and zm using (16),(17) and (19).
2. Calculate the values of m using (27).
3. Calculate the values of Pm using (24).
4. Set up the matrix A using (31),(32),(33) and (34) remembering to start with the 2(M + 1) 2(M + 1)
identity matrix.
5. Solve for Fm and Rm using (30).
16
6. Solve for the forward and reverse going waves at the interface using (25) where a1 and a2 in (25) are
Fm and Rm .
After solving for each frequency point, the waveforms can be solved using IDFT techniques.
1
Note that we have chosen here to view each section length as T = 2Fs
. This means that for each sample
taken, the waves actually traverse two sections. The reason for this is that because of how weve quantized
things, if we look at the front of the line, we will get non-zero reverse wave samples only when we do
things this way. As an example, on the first sample applied to the front of the line, we immediately see
the impedance discontinuity due to the first section. But, until the impulse has gone to the boundary of the
next section and back through the first section, we will not see that impedance discontinuity. Setting the
length as we do allows each sample at the front of the line to see every impedance discontinuity. Having
described the forward problem we now describe an algorithm to compute the and z for each section of
the transmission line from the two-port S-parameter measurements.
V0 |t=0 = Vf0 |t=0 (1 + 0 ) Z0. Again, knowing the forward wave in conjunction with the measured voltage
allows the calculation of the reflection coefficient at the immediate interface 0 . In the second sample,
we see the next forward wave along with a reflection at the interface to the next transmission line section.
Remember, we have construed our model so that on each sample, the wave traverses two delay elements.
Keeping to an impulsive forward wave, we see the wave that passed through the interface between the
source (which is 1 + 0 times our forward impulse) reflecting off the interface to the next section and
back again. Note that this wave reflects off the initial interface and some of this wave turns around and
becomes a forward propagating wave. As time continues things become much more complicated and the
complication escalates as time goes on as waves reflect off of interfaces along the line.
It is good to look at Figure 16 to gain some insight into the complexity, but fortunately, this insight is
not needed for the calculation of the impedance profile. All that is necessary is to know that the reflection
coefficient due to the first interface can be calculated directly.
Assume that we have a measurement of S11 looking into the port of the system. S11 is a frequencydomain measurement that provides, for each frequency, the returned, or reverse going standing wave or
phasor that comes back from the input port due to a unity incident or forward launched standing wave into
the port. If we launch an impulse into this system, we already know that the impulse reflected back at time
17
zero is 0 in size. But this is just the inverse DFT of S11 at time zero. Furthermore, since we are only
interested in the first point, we calculate 0 as:
[
]
N1
1
0 = IDFT (S11 )0 =
S110 + Re (S11N ) + 2
Re (S11n )
2N
n=1
(35)
Having now calculated for the first section, we create a small transmission line section and de-embed
it from the remainder of the system. The S-parameters for the ideal transmission line are given by (18)
and that can be used to de-embed the small section from the measured two-port system S-parameters Ssys .
De-embedding techniques described in [5] can be used to obtain the new system SsysNew . For convenience
the S-parameters are given below.
(
)
SL11 SL12
SL =
(36)
SL21 SL22
(
)
S11 S12
(37)
S=
S21 S22
SR =
S11 SL11
SL21 S12
SL12 S21 SL22 |S| S22 |SL |
SL22 S11 |SL |
(38)
Here SL is the two-port S-parameters of the ideal transmission line section obtained from and z (as
shown in (18)), S is the two-port S-parameters of the measured system, and SR is the two-port S-parameters
obtained by de-embedding the ideal transmission line section from the measured system. Also, |S| is the
determinant of the matrix (37) and |SL | is the determinant of the matrix (36).
Lets recapitulate what we know:
1. The IDFT of S11 of a system defines the reverse propagating waves emanating from a port after the
application of an impulse forward wave stimulus.
2. While the nature of this response is very complicated, we know that the very first value at time zero
is the of the first transmission line section.
3. If we know the of a section, and use a fixed time length for the section equal to one half of a sample
period,we can construct an s-parameter model of the first incremental transmission line section.
Incorporating non-ideal loss and group delay (as a function of frequency) is not addressed here,
but the loss can be estimated if S21 measurement of the system is available.
Having knowledge of the system (i.e. the S-parameters) and the S-parameters for the first transmission line
section provides a great advantage. This is because we can de-embed the first transmission line section
from the system and repeat the exercise to determine the next value of the second section. We can
therefore iteratively arrive at the values for the entire system.
Therefore, the method of generating the impedance profile can be performed by repeated calculating
the reflection coefficient of the first section in the remaining system, generating an s-parameter model of
a small transmission line section, de-embedding the section and doing this over and over again.
18
S11 SL11
S11 SL22 + SL12 SL21 SL11 SL22
(40)
Plugging in the S-parameters of the small section calculated from a given , we obtain the new S11 :
S11new =
S11 + S11 2 z2 z2 +
2 + S11 z2 S11 z2
(41)
The algorithm is summarized in Figure 19. Remember that we use (41) to calculate a new value of S11
for the remaining line after calculating a new value of for the impedance profile. If we did a Taylor series
expansion of S11new , we get:
(
)(
)
( )
1
S211
S11
1 + 2 + O 2 .
(42)
S11new = 2 + 1 2
z
z
z
If we neglect all powers of (implying a very small value of i.e. characteristic impedance has a value
close to Z0 ), then S11new is given as:
S11
.
(43)
z2
(43) says that for small values of , a good approximation of the S11 of the remaining line is simply
the S11 advanced (as opposed to delayed) by twice the length of the small transmission line section we are
using to approximate the line. This means that if we use a length that is one-half of a sample, the new S11
is the original S11 advanced by one sample. When we say a sample here, we mean a sample in the IDFT
of S11 . This means that each value of can be computed from the first sample of the IDFT of S11 which
means that we can determine each value of by simply sampling the IDFT of S11 . It turns out that under
most realistic circumstances, this is a very good and simple approximation of :
S11new
IDFT (S11 ) .
19
RealVector ImpedanceProfile(S11 , M)
Arguments:
ComplexVector S11 {
An N+1 element vector of S11 for each frequency
point including DC and the Nyquist rate
}
int M {
The number of sections desired for the profile
}
int m {0, , M 1}
]
[
N1
1
Re (S11n )
m 2N S110 + Re (S11N ) + 2
n=1
int n {0, . . . , N}
n
Complex z ej 2N
S11n
S11n +S11n 2m z2 m z2 +m
2m +S11n m z2 S11n m z2
return
Figure 19: Impedance Profile Calculation Algorithm
Simulation results are described next and will compare the performance of the peeling algorithm with
the IDFT approach.
Figure 20: Impedance profile for cable terminated with different terminations
Zc = 150
Td = 20 ps
Port 1
Zc = 50
Td = 1 ns
Zc = 165
Td = 20 ps
Zc = 50
Td = 500 ps
Port 2
Zc = 50
Td = 200 ps
Figure 23: Step response due to S21 for the DUT described in Figure 21.
comings in the proposed algorithm that we have not addressed in the paper. First, what happens when we
do not know the actual loss of the system? (41) that describes the calculation of new S11 needs and z from
the previous section. If G(f) and D(f) (gain and frequency dependent delay) are unknown functions, then
the z used in the algorithm will be an idealization as given in (17). This will introduce errors in peeling
equation (41). Second, we have not addressed the accuracy of peeling algorithm in the presence of noise.
In practice, any error in calculating or z for a section of transmission line (due to noise or idealization
of z), would introduce an error in the new S11 calculated using (41). This would cause the calculation of
subsequent to be more noisy and the cycle of error propagation would continue. One proposed solution
would be to use the knowledge of S21 to refine the value of and/or obtain a non-ideal value for z. Step
response due to S21 for the system shown in Figure 21 is shown in Figure 23. The effects of multiple
discontinuities in impedance and loss can be seen in this plot. The small blips that occur after the rising
edge can be attributed to the multiple reflections acting as secondary sources. Using this information (or
the actual S21 ) in refining the peeling algorithm would certainly result in a better algorithm.
22
Conclusions
We have demonstrated impedance computation problems caused by insufficient resolution and band limiting of the S-parameters. Based on our research, we recommend that all data-purification procedures,
including rational polynomial fitting, interpolation or extrapolation, should be applied to the raw nonconverted data rather than voltage transfer function or any similar function of S-parameters. By comparing
IDFT and RFE based approaches, we found that neither of them is perfect, but the best quality of solution
can be achieved by combining their strengths. By using both methods, we were able to cross-verify the
results and get a better judgment of the impedance profile computations.
We also provided an alternate algorithm that accounts for the multiple reflections due to impedance
discontinuities. We compared the new algorithm to the well-known IDFT approach and showed for what
kind of DUTs the IDFT algorithm can be used. Drawbacks of the proposed algorithm were described and
a possible solution was mentioned to further improve the accuracy of the proposed algorithm.
Acknowledgments
The authors wish to thank Balamurali Bhat, Aditi Sheth and Matthew Weinstein of Teledyne LeCroy for
providing the data to validate the peeling algorithm.
References
[1] I. Novak, Y. L. Li, E. Kunz, S. Paydavosi, L. Kocumbinski, K. Hinckley, A. Nosovitski, N. Shannon,
J. Miller, and G. Blando, Determining PCB Trace Impdedance by TDR: Challenges and Possible
Solutions, in DesignCon, UBM, February 2013.
[2] B. Gustavsen and A. Semylen, Rational Approximation of Frequency Domain Responses by Vector
Fitting, IEEE Trans. Power Delivery, vol. 14, pp. 10521061, July 2003.
[3] P. Pupalaikis, The Relationship Between Discrete Frequency S-parameters and Continuous Frequency
Responses, in DesignCon, UBM, February 2012.
[4] A. Semylen and B. Gustavsen, Vector Fitting by Pole Relocation for the State Equation Approximation of Non-rational Transfer Matrices, Circuit Systems Signal Processing, vol. 19, no. N6, 2000.
[5] P. Pupalaikis, K. Doshi, D. Dunham, and D. DeGroot, De-embedding in High Speed Design, in
DesignCon, UBM, February 2012.
23